Planet AI Weekly: Planet AI Weekly July 19, 2026

This week we reviewed 256 articles from 20 sources.

Moonshot AI released Kimi K3, a 2.8T-parameter open MoE model under MIT license with 1M-token context, undercutting proprietary models on cost per token in early benchmarks. OpenAI's hardware ambitions surfaced via a screenless mobile speaker, while Apple's lawsuit against OpenAI raises stakes on training data and partnerships. The central tension this week: agents ace technical evaluations but fail financial ones, as engineering success collides with CFO spreadsheets.

Official Highlights

Moonshot AI ships Kimi K3, a 2.8T-parameter open MoE with 1M-token context, activating 16 of 896 experts. MIT licensing and measured cost per token undercut proprietary alternatives on enterprise workloads. Read more

NVIDIA ships DeepStream 9.1 with 13 agentic skills that let coding agents build multi-camera video analytics pipelines from natural-language prompts. Multi-View 3D Tracking fuses detections into persistent 3D trajectories for real-time enterprise surveillance. Read more

AWS makes OpenAI's GPT-5.6 Sol, Terra, and Luna generally available on Bedrock, bringing the models to AWS's next-gen inference engine. The integration targets enterprise agentic workloads requiring high-throughput, low-latency inference. Read more

OpenAI's first hardware device is reportedly a screenless speaker that can move. The form factor signals a strategic shift toward ambient, physical agent interfaces. Read more

Siri AI is becoming Apple's everything tool. The revamped iOS 27 beta integrates Siri as system-wide backbone, routing on-device AI across the iPhone experience with privacy-first architecture. Read more

From the Community

Your AI agent passed every eval. Finance still killed it. Pratik Rupareliya reveals the one metric that predicts agent survival in production: cost per resolution, with a tested framework for measuring ROI before deployment. Read more

54% of enterprises have had an AI agent incident, and most still let agents share credentials. A VentureBeat analysis of 107 orgs shows agent security controls lag behind adoption, with credential-sharing as the top vulnerability. Read more

Loop engineering with adaptive PDF parsing: start cheap, pay for a heavier parser only when the page needs it. Angela Shi's guide details a free, deterministic escalation cascade that flags failed parses before invoking costly models. Read more

Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2: open trillion-scale MoE models compared on benchmarks, license, and serving cost. Michal Sutter's breakdown shows Kimi K3's MIT license and lower serving costs give it an edge for enterprise adoption. Read more

Agentic orchestration: enterprise AI orgs have a deployment problem, not a platform problem. A VentureBeat survey of 101 enterprises finds Anthropic's Claude leads orchestration, but most teams mislabel chatbots as agents, masking execution gaps. Read more

OpenAI's GPT-Red, an automated red-teaming model, beat human red-teamers 84% to 13% on prompt injection. The internal model uses self-play reinforcement learning and discovered a novel "Fake Compliance" attack vector. Read more

Could your AI systems already be high-risk under the EU AI Act? This on-demand webinar breaks down the latest guidance and what enterprise governance programs must do next. Read more

Satya Nadella warned enterprises about proprietary model risks in a rare public post. The Microsoft CEO cited vendor lock-in, cost unpredictability, and lack of transparency, positioning open-weight alternatives as safer long-term bets. Read more

Featured This Week

"Your AI Agent Passed Every Eval. Finance Still Killed It" by Pratik Rupareliya reframes agent deployment around cost per resolution. The post dissects why agents that excel in technical evaluations fail in production, where their "successful" outcomes cost more than human labor. Rupareliya provides a tested framework for measuring ROI before launch, with methods for calculating break-even thresholds. For engineering leaders deploying agents, this bridges the gap between eval harnesses and CFO approval. Read more

Send comments and story tips to tips@planet-ai.net. If you found this useful, share it with a colleague.